Summary
Downloading the on-device image generation models (Bonsai Image 4B, FLUX.2 klein,
Z-Image Turbo) is effectively unusable for users in mainland China: the download runs
from huggingface.co / GitHub, which is heavily throttled here. Even with a VPN active
the transfer stalls (e.g. a single .gallerytmp stuck at ~110 KB). It would help a lot
to let users either point Box at a mirror, or install model files manually/offline.
Environment
- Device: Redmi K80 (Snapdragon 8 Gen 3), 12/16 GB RAM
- Android: 15
- Box version: v3.4.5 (from GitHub releases)
- Model: Bonsai Image 4B (~4.3 GB)
- Network: mainland China; VPN on; download still stalls
Steps to reproduce
- Open Box → Image Gen → Bonsai Image 4B → Download.
- Watch the progress.
- Transfer stalls almost immediately.
Actual behaviour
- Exactly one temp file is created, e.g.
Android/data/com.box.gallery/files/Bonsai_Image_4B/main/dit_gpu_int4b32.tflite.gallerytmp
- Its size stays at ~110 KB; the download never progresses.
- Setting
Range/resume works in code, but the source is too slow to make progress.
Expected behaviour
- Download completes in reasonable time, OR
- A way to obtain the model without the built-in downloader.
Proposal (any of these would help)
- Configurable download base URL / mirror. The image models are hosted at
huggingface.co/litert-community/... and there is a well-known, much faster mirror at
hf-mirror.com. A setting to replace the download host (or an "HF_ENDPOINT" style
option) would fix this for everyone in CN.
- Offline / manual install path. Allow the user to drop the official files into the
model folder (e.g. .../files/<Model>/<version>/) and have Box recognise them, instead
of requiring the in-app download to reach SUCCEEDED. The files are public at
litert-community/Bonsai-Image-ternary-4B, litert-community/FLUX.2-klein-4B-LiteRT
and litert-community/Z-Image-Turbo-LiteRT.
- Optionally, resume-friendly mirror fallback: if the primary host stalls, retry via
the mirror automatically.
Why this matters
Box is one of the very few apps that can run these LiteRT image models on-device, but for
a large user base the model download is the one hard blocker. Fixing the download path
would unlock the whole image-generation feature for CN users.
Thanks for the great project!
Summary
Downloading the on-device image generation models (Bonsai Image 4B, FLUX.2 klein,
Z-Image Turbo) is effectively unusable for users in mainland China: the download runs
from huggingface.co / GitHub, which is heavily throttled here. Even with a VPN active
the transfer stalls (e.g. a single
.gallerytmpstuck at ~110 KB). It would help a lotto let users either point Box at a mirror, or install model files manually/offline.
Environment
Steps to reproduce
Actual behaviour
Android/data/com.box.gallery/files/Bonsai_Image_4B/main/dit_gpu_int4b32.tflite.gallerytmpRange/resume works in code, but the source is too slow to make progress.Expected behaviour
Proposal (any of these would help)
huggingface.co/litert-community/...and there is a well-known, much faster mirror athf-mirror.com. A setting to replace the download host (or an "HF_ENDPOINT" styleoption) would fix this for everyone in CN.
model folder (e.g.
.../files/<Model>/<version>/) and have Box recognise them, insteadof requiring the in-app download to reach
SUCCEEDED. The files are public atlitert-community/Bonsai-Image-ternary-4B,litert-community/FLUX.2-klein-4B-LiteRTand
litert-community/Z-Image-Turbo-LiteRT.the mirror automatically.
Why this matters
Box is one of the very few apps that can run these LiteRT image models on-device, but for
a large user base the model download is the one hard blocker. Fixing the download path
would unlock the whole image-generation feature for CN users.
Thanks for the great project!